Bookend every AI session with a plan and debrief

The integrity risk in AI-assisted study is drift: a session that starts as revision quietly becomes the model doing the work. Bookends prevent it. Before opening the tool, the student writes a specific goal and a confidence rating; afterwards, they write what changed in their understanding and what they did versus what the AI did. Zimmerman (2000) puts this plan, monitor, reflect cycle at the heart of self-regulated learning, and it is precisely the cycle an eager assistant erodes. (Manning's srl-session-wrapper skill scripts all three checkpoints.)

In practice

Before you open the AI:
  Goal (specific, checkable): ...
  Confidence I can do this now (0-100): ...
About 15 minutes in:
  Is this approach working? What changes?
  Who is doing the thinking right now?
When you close it:
  What changed in my understanding?
  What did I do, and what did the AI do?
  What could I now do without the AI?

• The opening goal must be checkable: 'be able to derive the quadratic formula unaided', never 'understand quadratics better'.
• Add one mid-session pause around fifteen minutes in: is this approach working, and who is doing the thinking right now?
• The debrief names ownership explicitly: what I did, what the AI did, and what I could now do with the AI switched off.
• Collect the bookends rather than chat transcripts; the debrief extends the usage log idea (see "Have students log what the AI did for them").